现在将# 将 4输入分开,构建新的相同模态结合的2输入,2分支 import math import logging from functools import partial…
- User
现在将# 将 4输入分开,构建新的相同模态结合的2输入,2分支 import math import logging from functools import partial from collections import OrderedDict from copy import deepcopy
import torch import torch.nn as nn import torch.nn.functional as F
from timm.models.layers import to_2tuple
from lib.models.layers.patch_embed import PatchEmbed, PatchEmbed_event, xcorr_depthwise from .utils import combine_tokens, recover_tokens from .vit import VisionTransformer from ..layers.attn_blocks import CEBlock
_logger = logging.getLogger(name)
class VisionTransformerCE(VisionTransformer): """ Vision Transformer with candidate elimination (CE) module
A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale` - https://arxiv.org/abs/2010.11929 Includes distillation token & head support for `DeiT: Data-efficient Image Transformers` - https://arxiv.org/abs/2012.12877 """ def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, representation_size=None, distilled=False, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., embed_layer=PatchEmbed, norm_layer=None, act_layer=None, weight_init='', ce_loc=None, ce_keep_ratio=None): super().__init__() if isinstance(img_size, tuple): self.img_size = img_size else: self.img_size = to_2tuple(img_size) self.patch_size = patch_size self.in_chans = in_chans self.num_classes = num_classes self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models self.num_tokens = 2 if distilled else 1 norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6) act_layer = act_layer or nn.GELU self.patch_embed = embed_layer( img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim) num_patches = self.patch_embed.num_patches self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) self.dist_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if distilled else None self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim)) self.pos_drop = nn.Dropout(p=drop_rate) self.pos_embed_event = PatchEmbed_event(in_chans=32, embed_dim=768, kernel_size=4, stride=4) dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule blocks = [] ce_index = 0 self.ce_loc = ce_loc for i in range(depth): ce_keep_ratio_i = 1.0 if ce_loc is not None and i in ce_loc: ce_keep_ratio_i = ce_keep_ratio[ce_index] ce_index += 1 blocks.append( CEBlock( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, act_layer=act_layer, keep_ratio_search=ce_keep_ratio_i) ) self.blocks = nn.Sequential(*blocks) self.norm = norm_layer(embed_dim) self.init_weights(weight_init) def forward_features(self, z, x, event_z, event_x, mask_z=None, mask_x=None, ce_template_mask=None, ce_keep_rate=None, return_last_attn=False ): # 分支1 处理流程 B, H, W = x.shape[0], x.shape[2], x.shape[3] x = self.patch_embed(x) z = self.patch_embed(z) z += self.pos_embed_z x += self.pos_embed_x if mask_z is not None and mask_x is not None: mask_z = F.interpolate(mask_z[None].float(), scale_factor=1. / self.patch_size).to(torch.bool)[0] mask_z = mask_z.flatten(1).unsqueeze(-1) mask_x = F.interpolate(mask_x[None].float(), scale_factor=1. / self.patch_size).to(torch.bool)[0] mask_x = mask_x.flatten(1).unsqueeze(-1) mask_x = combine_tokens(mask_z, mask_x, mode=self.cat_mode) mask_x = mask_x.squeeze(-1) if self.add_cls_token: cls_tokens = self.cls_token.expand(B, -1, -1) cls_tokens = cls_tokens + self.cls_pos_embed if self.add_sep_seg: x += self.search_segment_pos_embed z += self.template_segment_pos_embed x = combine_tokens(z, x, mode=self.cat_mode) if self.add_cls_token: x = torch.cat([cls_tokens, x], dim=1) x = self.pos_drop(x) lens_z = self.pos_embed_z.shape[1] lens_x = self.pos_embed_x.shape[1] global_index_t = torch.linspace(0, lens_z - 1, lens_z).to(x.device) global_index_t = global_index_t.repeat(B, 1) global_index_s = torch.linspace(0, lens_x - 1, lens_x).to(x.device) global_index_s = global_index_s.repeat(B, 1) removed_indexes_s = [] for i, blk in enumerate(self.blocks): x, global_index_t, global_index_s, removed_index_s, attn = \ blk(x, global_index_t, global_index_s, mask_x, ce_template_mask, ce_keep_rate) if self.ce_loc is not None and i in self.ce_loc: removed_indexes_s.append(removed_index_s) x = self.norm(x) # # [bs, n_patch, dim] = [bs, 320, 768] 320 = 64 + 256 # # 分支2 处理流程 event_x = self.pos_embed_event(event_x) event_z = self.pos_embed_event(event_z) event_x += self.pos_embed_x event_z += self.pos_embed_z event_x = combine_tokens(event_z, event_x, mode=self.cat_mode) if self.add_cls_token: event_x = torch.cat([cls_tokens, event_x], dim=1) lens_z = self.pos_embed_z.shape[1] lens_x = self.pos_embed_x.shape[1] global_index_t1 = torch.linspace(0, lens_z - 1, lens_z).to(event_x.device) global_index_t1 = global_index_t1.repeat(B, 1) global_index_s1 = torch.linspace(0, lens_x - 1, lens_x).to(event_x.device) global_index_s1 = global_index_s1.repeat(B, 1) removed_indexes_s1 = [] for i, blk in enumerate(self.blocks): event_x, global_index_t1, global_index_s1, removed_index_s1, attn = \ blk(event_x, global_index_t1, global_index_s1, mask_x, ce_template_mask, ce_keep_rate) if self.ce_loc is not None and i in self.ce_loc: removed_indexes_s1.append(removed_index_s1) event_x = self.norm(event_x) # print('x.shape: ',x.shape) #x.shape: torch.Size([2, 320, 768]) # print('event_x.shape: ',event_x.shape) # event_x.shape: torch.Size([2, 320, 768]) x_cat = torch.cat([x, event_x], dim=1) # print('x_cat.shape: ',x_cat.shape) # x_cat.shape: torch.Size([2, 640, 768]) x = x_cat aux_dict = { "attn": attn, "removed_indexes_s": removed_indexes_s, # used for visualization } return x, aux_dict def forward(self, z, x, event_z, event_x, ce_template_mask=None, ce_keep_rate=None, tnc_keep_rate=None, return_last_attn=False): x, aux_dict = self.forward_features(z, x, event_z, event_x, ce_template_mask=ce_template_mask, ce_keep_rate=ce_keep_rate,) return x, aux_dictdef _create_vision_transformer(pretrained=False, **kwargs): model = VisionTransformerCE(**kwargs)
if pretrained: if 'npz' in pretrained: model.load_pretrained(pretrained, prefix='') else: checkpoint = torch.load(pretrained, map_location="cpu") missing_keys, unexpected_keys = model.load_state_dict(checkpoint["model"], strict=False) print('Load pretrained model from: ' + pretrained) return modeldef vit_base_patch16_224_ce(pretrained=False, **kwargs): """ ViT-Base model (ViT-B/16) from original paper (https://arxiv.org/abs/2010.11929). """ model_kwargs = dict( patch_size=16, embed_dim=768, depth=12, num_heads=12, **kwargs) model = _create_vision_transformer(pretrained=pretrained, **model_kwargs) return model
def vit_large_patch16_224_ce(pretrained=False, **kwargs): """ ViT-Large model (ViT-L/16) from original paper (https://arxiv.org/abs/2010.11929). """ model_kwargs = dict( patch_size=16, embed_dim=1024, depth=24, num_heads=16, **kwargs) model = create_vision_transformer(pretrained=pretrained, **model_kwargs) return model中event部分的特征提取网络改为import torch import torch.nn as nn from spikingjelly.clock_driven.neuron import MultiStepParametricLIFNode, MultiStepLIFNode from spikingjelly.clock_driven import layer from timm.models.layers import to_2tuple, trunc_normal, DropPath from timm.models.registry import register_model from timm.models.vision_transformer import _cfg from einops.layers.torch import Rearrange import torch.nn.functional as F from functools import partial from timm.models import create_model import pdb from Bottleneck import BottleneckTransformer
from mmengine.model.weight_init import kaiming_init
from mmcv.cnn import kaiming_init
all = ['Spikingformer'] tau_thr = 1.5
class MLP(nn.Module): def init(self, in_features, hidden_features=None, out_features=None, drop=0.): super().init() out_features = out_features or in_features hidden_features = hidden_features or in_features self.mlp1_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) #, backend='cupy' self.mlp1_conv = nn.Conv2d(in_features, hidden_features, kernel_size=1, stride=1) self.mlp1_bn = nn.BatchNorm2d(hidden_features)
self.mlp2_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.mlp2_conv = nn.Conv2d(hidden_features, out_features, kernel_size=1, stride=1) self.mlp2_bn = nn.BatchNorm2d(out_features) self.c_hidden = hidden_features self.c_output = out_features def forward(self, x): T, B, C, H, W = x.shape x = self.mlp1_lif(x) x = self.mlp1_conv(x.flatten(0, 1)) x = self.mlp1_bn(x).reshape(T, B, self.c_hidden, H, W) x = self.mlp2_lif(x) x = self.mlp2_conv(x.flatten(0, 1)) x = self.mlp2_bn(x).reshape(T, B, C, H, W) return xclass SpikingSelfAttention(nn.Module): def init(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1): super().init() assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}."
self.dim = dim self.num_heads = num_heads self.proj_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.q_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.q_bn = nn.BatchNorm1d(dim) self.q_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.k_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.k_bn = nn.BatchNorm1d(dim) self.k_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.v_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.v_bn = nn.BatchNorm1d(dim) self.v_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.attn_lif = MultiStepLIFNode(tau=tau_thr, v_threshold=0.5, detach_reset=True) self.proj_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1) self.proj_bn = nn.BatchNorm1d(dim) def forward(self, x): T, B, C, H, W = x.shape x = self.proj_lif(x) x = x.flatten(3) T, B, C, N = x.shape x_for_qkv = x.flatten(0, 1) q_conv_out = self.q_conv(x_for_qkv) q_conv_out = self.q_bn(q_conv_out).reshape(T, B, C, N) q_conv_out = self.q_lif(q_conv_out) q = q_conv_out.transpose(-1, -2).reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4) #torch.Size([4, 2, 16, 336, 16]) k_conv_out = self.k_conv(x_for_qkv) k_conv_out = self.k_bn(k_conv_out).reshape(T, B, C, N) k_conv_out = self.k_lif(k_conv_out) k = k_conv_out.transpose(-1, -2).reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4)#torch.Size([4, 2, 16, 336, 16]) v_conv_out = self.v_conv(x_for_qkv) v_conv_out = self.v_bn(v_conv_out).reshape(T, B, C, N) v_conv_out = self.v_lif(v_conv_out) v = v_conv_out.transpose(-1, -2).reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4) #torch.Size([4, 2, 16, 336, 16]) attn = (q @ k.transpose(-2, -1)) #torch.Size([4, 2, 16, 336, 336]) x = (attn @ v) * 0.125 #torch.Size([4, 2, 16, 336, 16]) x = x.transpose(3, 4).reshape(T, B, C, N) x = self.attn_lif(x) x = x.flatten(0, 1) x = self.proj_bn(self.proj_conv(x)).reshape(T, B, C, H, W) #torch.Size([4, 2, 256, 21, 16]) return xclass SpikingTransformer(nn.Module): def init(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., norm_layer=nn.LayerNorm, sr_ratio=1): super().init() self.norm1 = norm_layer(dim) self.attn = SpikingSelfAttention(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = MLP(in_features=dim, hidden_features=mlp_hidden_dim, drop=drop)
def forward(self, x): x = x + self.attn(x) x = x + self.mlp(x) return xclass SpikingTokenizer(nn.Module): def init(self, img_size_h=128, img_size_w=128, patch_size=4, in_channels=2, embed_dims=256): super().init() self.image_size = [img_size_h, img_size_w] patch_size = to_2tuple(patch_size) self.patch_size = patch_size self.C = in_channels self.H, self.W = self.image_size[0] // patch_size[0], self.image_size[1] // patch_size[1] self.num_patches = self.H * self.W
self.proj_conv = nn.Conv2d(in_channels, embed_dims//8, kernel_size=3, stride=1, padding=1, bias=False) self.proj_bn = nn.BatchNorm2d(embed_dims//8) self.proj1_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.maxpool1 = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) self.proj1_conv = nn.Conv2d(embed_dims//8, embed_dims//4, kernel_size=3, stride=1, padding=1, bias=False) self.proj1_bn = nn.BatchNorm2d(embed_dims//4) self.proj2_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.maxpool2 = torch.nn.MaxPool2d(kernel_size=(3,2), stride=2, padding=0, dilation=1, ceil_mode=False) self.proj2_conv = nn.Conv2d(embed_dims//4, embed_dims//2, kernel_size=3, stride=1, padding=1, bias=False) self.proj2_bn = nn.BatchNorm2d(embed_dims//2) self.proj3_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.maxpool3 = torch.nn.MaxPool2d(kernel_size=(2,3), stride=2, padding=0, dilation=1, ceil_mode=False) self.proj3_conv = nn.Conv2d(embed_dims//2, embed_dims, kernel_size=3, stride=1, padding=1, bias=False) self.proj3_bn = nn.BatchNorm2d(embed_dims) self.proj4_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.maxpool4 = torch.nn.MaxPool2d(kernel_size=(3,2), stride=2, padding=0, dilation=1, ceil_mode=False) self.proj4_conv = nn.Conv2d(embed_dims, embed_dims, kernel_size=3, stride=1, padding=1, bias=False) self.proj4_bn = nn.BatchNorm2d(embed_dims) def forward(self, x): T, B, C, H, W = x.shape x = self.proj_conv(x.flatten(0, 1)) x = self.proj_bn(x).reshape(T,B,-1,H,W).contiguous() x = self.proj1_lif(x).flatten(0,1).contiguous() x = self.maxpool1(x) x = self.proj1_conv(x) x = self.proj1_bn(x).reshape(T, B, -1, H//2, W//2).contiguous() x = self.proj2_lif(x).flatten(0, 1).contiguous() x = self.maxpool2(x) x = self.proj2_conv(x) x = self.proj2_bn(x).reshape(T, B, -1, H//4, W//4).contiguous() x = self.proj3_lif(x).flatten(0, 1).contiguous() x = self.maxpool3(x) x = self.proj3_conv(x) x = self.proj3_bn(x).reshape(T, B, -1, H//8, W//8).contiguous() x = self.proj4_lif(x).flatten(0, 1).contiguous() x = self.maxpool4(x) x = self.proj4_conv(x) x = self.proj4_bn(x).reshape(T, B, -1, H//16, W//16).contiguous() return x, (None, None)class vit_snn(nn.Module): def init(self, img_size_h=128, img_size_w=128, patch_size=16, in_channels=2, num_classes=11, embed_dims=[64, 128, 256], num_heads=[1, 2, 4], mlp_ratios=[4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, depths=[6, 8, 6], sr_ratios=[8, 4, 2], T=4, pretrained_cfg=None, ): super().init() self.num_classes = num_classes self.depths = depths self.T = T dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depths)]
patch_embed = SpikingTokenizer(img_size_h=img_size_h, img_size_w=img_size_w, patch_size=patch_size, in_channels=in_channels, embed_dims=embed_dims) num_patches = patch_embed.num_patches block = nn.ModuleList([SpikingTransformer( dim=embed_dims, num_heads=num_heads, mlp_ratio=mlp_ratios, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[j], norm_layer=norm_layer, sr_ratio=sr_ratios) for j in range(depths)]) self.Bottleneck_block = nn.ModuleList([BottleneckTransformer( dim=embed_dims, num_heads=4, mlp_ratio=mlp_ratios, qkv_bias=qkv_bias, drop=0., attn_drop=0., init_values=None, drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm) for j in range(2)]) setattr(self, f"patch_embed", patch_embed) setattr(self, f"block", block) # classification head self.head = nn.Linear(8192, num_classes) if num_classes > 0 else nn.Identity() self.apply(self._init_weights) self.bottleneck_initfeature = nn.Parameter(torch.randn(1, 64, 256)) kaiming_init(self.bottleneck_initfeature, mode='fan_out') self.pos_embed = nn.Parameter(torch.zeros(1, 400, 256)) self.pos_drop = nn.Dropout(p=0.) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def forward_features(self, x): block = getattr(self, f"block") patch_embed = getattr(self, f"patch_embed") x, (H, W) = patch_embed(x) #x--torch.Size([4, 2, 256, 21, 16]) for blk in block: x = blk(x) #torch.Size([16, 8, 256, 21, 16]) #torch.Size([4, 2, 256, 21, 16]) T,B,C,H,W =x.shape botlk_x = x.flatten(3) #torch.Size([4, 2, 256, 336]) T, B, C, N botlk_x = botlk_x.flatten(0, 1).permute(0,2,1) #torch.Size([8, 336, 256]) bottleneck_feature = self.bottleneck_initfeature.expand(T*B,64,C) fusion_feature_agg = torch.cat((bottleneck_feature,botlk_x),1) #BT, 672,256 fusion_feature_agg = self.pos_drop(fusion_feature_agg + self.pos_embed) # for bottleneck_blk in self.Bottleneck_block: fusion_feature_agg = bottleneck_blk(fusion_feature_agg) bottleneck_feature = fusion_feature_agg[:,0:336,:].mean(1).reshape(B,1,-1) #torch.Size([16, 2, 336, 256]) return x.flatten(3).mean(3), bottleneck_feature #torch.Size([4, 2, 256]) def forward(self, event_tensor): B, N, C, H, W = event_tensor.shape event_tensor = event_tensor.permute(1, 0, 2, 3, 4) # [T, N, 2, *, *] torch.Size([16, 2, 12, 346, 260]) event_tensor, bottleneck_feature = self.forward_features(event_tensor) #torch.Size([16, 2, 256, 21, 16]) torch.Size([16, 2, 336, 256]) event_tensor = event_tensor.permute(1,0,2).reshape(B,-1) x1 = self.head(event_tensor) return x1@register_model def Spikingformer(pretrained=False, **kwargs): model = vit_snn( patch_size=16, embed_dims=256, num_heads=8, mlp_ratios=4, in_channels=12, num_classes=300, qkv_bias=False, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=2, sr_ratios=1, **kwargs ) model.default_cfg = _cfg() return model
。
- Assistant
好的,让我们将 VisionTransformerCE 类中的 event 部分的特征提取网络替换为 vit_snn。在这个过程中,我们需要调整相关代码,以确保新特征提取器的输入和输出满足 VisionTransformerCE 的需求。以下是更新后的代码:
import math import logging from functools import partial from collections import OrderedDict from copy import deepcopy import torch import torch.nn as nn import torch.nn.functional as F from timm.models.layers import to_2tuple from lib.models.layers.patch_embed import PatchEmbed, xcorr_depthwise from .utils import combine_tokens, recover_tokens from .vit import VisionTransformer from ..layers.attn_blocks import CEBlock # Spikingformer related imports import torch import torch.nn as nn from spikingjelly.clock_driven.neuron import MultiStepParametricLIFNode, MultiStepLIFNode from spikingjelly.clock_driven import layer from timm.models.layers import to_2tuple, trunc_normal_, DropPath from timm.models.registry import register_model from timm.models.vision_transformer import _cfg from einops.layers.torch import Rearrange import torch.nn.functional as F from functools import partial from timm.models import create_model from Bottleneck import BottleneckTransformer from mmcv.cnn import kaiming_init _logger = logging.getLogger(__name__) class VisionTransformerCE(VisionTransformer): """ Vision Transformer with candidate elimination (CE) module A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale` - https://arxiv.org/abs/2010.11929 Includes distillation token & head support for `DeiT: Data-efficient Image Transformers` - https://arxiv.org/abs/2012.12877 """ def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, representation_size=None, distilled=False, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., embed_layer=PatchEmbed, norm_layer=None, act_layer=None, weight_init='', ce_loc=None, ce_keep_ratio=None): super().__init__() if isinstance(img_size, tuple): self.img_size = img_size else: self.img_size = to_2tuple(img_size) self.patch_size = patch_size self.in_chans = in_chans self.num_classes = num_classes self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models self.num_tokens = 2 if distilled else 1 norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6) act_layer = act_layer or nn.GELU self.patch_embed = embed_layer( img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim) num_patches = self.patch_embed.num_patches self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) self.dist_token = nn.Parameter(torch.zeros(1, 1, embed_dim)) if distilled else None self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + self.num_tokens, embed_dim)) self.pos_drop = nn.Dropout(p=drop_rate) self.pos_embed_event = vit_snn(img_size_h=img_size, img_size_w=img_size, patch_size=4, in_channels=32, embed_dims=768) # Replace with vit_snn dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule blocks = [] ce_index = 0 self.ce_loc = ce_loc for i in range(depth): ce_keep_ratio_i = 1.0 if ce_loc is not None and i in ce_loc: ce_keep_ratio_i = ce_keep_ratio[ce_index] ce_index += 1 blocks.append( CEBlock( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, act_layer=act_layer, keep_ratio_search=ce_keep_ratio_i) ) self.blocks = nn.Sequential(*blocks) self.norm = norm_layer(embed_dim) self.init_weights(weight_init) def forward_features(self, z, x, event_z, event_x, mask_z=None, mask_x=None, ce_template_mask=None, ce_keep_rate=None, return_last_attn=False ): # 分支1 处理流程 B, H, W = x.shape[0], x.shape[2], x.shape[3] x = self.patch_embed(x) z = self.patch_embed(z) z += self.pos_embed_z x += self.pos_embed_x if mask_z is not None and mask_x is not None: mask_z = F.interpolate(mask_z[None].float(), scale_factor=1. / self.patch_size).to(torch.bool)[0] mask_z = mask_z.flatten(1).unsqueeze(-1) mask_x = F.interpolate(mask_x[None].float(), scale_factor=1. / self.patch_size).to(torch.bool)[0] mask_x = mask_x.flatten(1).unsqueeze(-1) mask_x = combine_tokens(mask_z, mask_x, mode=self.cat_mode) mask_x = mask_x.squeeze(-1) if self.add_cls_token: cls_tokens = self.cls_token.expand(B, -1, -1) cls_tokens = cls_tokens + self.cls_pos_embed if self.add_sep_seg: x += self.search_segment_pos_embed z += self.template_segment_pos_embed x = combine_tokens(z, x, mode=self.cat_mode) if self.add_cls_token: x = torch.cat([cls_tokens, x], dim=1) x = self.pos_drop(x) lens_z = self.pos_embed_z.shape[1] lens_x = self.pos_embed_x.shape[1] global_index_t = torch.linspace(0, lens_z - 1, lens_z).to(x.device) global_index_t = global_index_t.repeat(B, 1) global_index_s = torch.linspace(0, lens_x - 1, lens_x).to(x.device) global_index_s = global_index_s.repeat(B, 1) removed_indexes_s = [] for i, blk in enumerate(self.blocks): x, global_index_t, global_index_s, removed_index_s, attn = \ blk(x, global_index_t, global_index_s, mask_x, ce_template_mask, ce_keep_rate) if self.ce_loc is not None and i in self.ce_loc: removed_indexes_s.append(removed_index_s) x = self.norm(x) # # [bs, n_patch, dim] = [bs, 320, 768] 320 = 64 + 256 # # 分支2 处理流程 event_x = self.pos_embed_event(event_x) # 使用自定义Spikingformer特征提取器 event_z = self.pos_embed_event(event_z) # 使用自定义Spikingformer特征提取器 event_x += self.pos_embed_x event_z += self.pos_embed_z event_x = combine_tokens(event_z, event_x, mode=self.cat_mode) if self.add_cls_token: event_x = torch.cat([cls_tokens, event_x], dim=1) lens_z = self.pos_embed_z.shape[1] lens_x = self.pos_embed_x.shape[1] global_index_t1 = torch.linspace(0, lens_z - 1, lens_z).to(event_x.device) global_index_t1 = global_index_t1.repeat(B, 1) global_index_s1 = torch.linspace(0, lens_x - 1, lens_x).to(event_x.device) global_index_s1 = global_index_s1.repeat(B, 1) removed_indexes_s1 = [] for i, blk in enumerate(self.blocks): event_x, global_index_t1, global_index_s1, removed_index_s1, attn = \ blk(event_x, global_index_t1, global_index_s1, mask_x, ce_template_mask, ce_keep_rate) if self.ce_loc is not None and i in self.ce_loc: removed_indexes_s1.append(removed_index_s1) event_x = self.norm(event_x) x_cat = torch.cat([x, event_x], dim=1) x = x_cat aux_dict = { "attn": attn, "removed_indexes_s": removed_indexes_s, # used for visualization } return x, aux_dict def forward(self, z, x, event_z, event_x, ce_template_mask=None, ce_keep_rate=None, tnc_keep_rate=None, return_last_attn=False): x, aux_dict = self.forward_features(z, x, event_z, event_x, ce_template_mask=ce_template_mask, ce_keep_rate=ce_keep_rate,) return x, aux_dict def _create_vision_transformer(pretrained=False, **kwargs): model = VisionTransformerCE(**kwargs) if pretrained: if 'npz' in pretrained: model.load_pretrained(pretrained, prefix='') else: checkpoint = torch.load(pretrained, map_location="cpu") missing_keys, unexpected_keys = model.load_state_dict(checkpoint["model"], strict=False) print('Load pretrained model from: ' + pretrained) return model def vit_base_patch16_224_ce(pretrained=False, **kwargs): """ ViT-Base model (ViT-B/16) from original paper (https://arxiv.org/abs/2010.11929). """ model_kwargs = dict( patch_size=16, embed_dim=768, depth=12, num_heads=12, **kwargs) model = _create_vision_transformer(pretrained=pretrained, **model_kwargs) return model def vit_large_patch16_224_ce(pretrained=False, **kwargs): """ ViT-Large model (ViT-L/16) from original paper (https://arxiv.org/abs/2010.11929). """ model_kwargs = dict( patch_size=16, embed_dim=1024, depth=24, num_heads=16, **kwargs) model = _create_vision_transformer(pretrained=pretrained, **model_kwargs) return model # Spikingformer definition tau_thr = 1.5 class MLP(nn.Module): def __init__(self, in_features, hidden_features=None, out_features=None, drop=0.): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.mlp1_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) #, backend='cupy' self.mlp1_conv = nn.Conv2d(in_features, hidden_features, kernel_size=1, stride=1) self.mlp1_bn = nn.BatchNorm2d(hidden_features) self.mlp2_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.mlp2_conv = nn.Conv2d(hidden_features, out_features, kernel_size=1, stride=1) self.mlp2_bn = nn.BatchNorm2d(out_features) self.c_hidden = hidden_features self.c_output = out_features def forward(self, x): T, B, C, H, W = x.shape x = self.mlp1_lif(x) x = self.mlp1_conv(x.flatten(0, 1)) x = self.mlp1_bn(x).reshape(T, B, self.c_hidden, H, W) x = self.mlp2_lif(x) x = self.mlp2_conv(x.flatten(0, 1)) x = self.mlp2_bn(x).reshape(T, B, C, H, W) return x class SpikingSelfAttention(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1): super().__init__() assert dim % num_heads == 0, f"dim {dim} should be divided by num_heads {num_heads}." self.dim = dim self.num_heads = num_heads self.proj_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.q_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.q_bn = nn.BatchNorm1d(dim) self.q_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.k_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.k_bn = nn.BatchNorm1d(dim) self.k_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.v_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1, bias=False) self.v_bn = nn.BatchNorm1d(dim) self.v_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.attn_lif = MultiStepLIFNode(tau=tau_thr, v_threshold=0.5, detach_reset=True) self.proj_conv = nn.Conv1d(dim, dim, kernel_size=1, stride=1) self.proj_bn = nn.BatchNorm1d(dim) def forward(self, x): T, B, C, H, W = x.shape x = self.proj_lif(x) x = x.flatten(3) T, B, C, N = x.shape x_for_qkv = x.flatten(0, 1) q_conv_out = self.q_conv(x_for_qkv) q_conv_out = self.q_bn(q_conv_out).reshape(T, B, C, N) q_conv_out = self.q_lif(q_conv_out) q = q_conv_out.transpose(-1, -2).reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4) #torch.Size([4, 2, 16, 336, 16]) k_conv_out = self.k_conv(x_for_qkv) k_conv_out = self.k_bn(k_conv_out).reshape(T, B, C, N) k_conv_out = self.k_lif(k_conv_out) k = k_conv_out.transpose(-1, -2).reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4)#torch.Size([4, 2, 16, 336, 16]) v_conv_out = self.v_conv(x_for_qkv) v_conv_out = self.v_bn(v_conv_out).reshape(T, B, C, N) v_conv_out = self.v_lif(v_conv_out) v = v_conv_out.transpose(-1, -2).reshape(T, B, N, self.num_heads, C // self.num_heads).permute(0, 1, 3, 2, 4) #torch.Size([4, 2, 16, 336, 16]) attn = (q @ k.transpose(-2, -1)) #torch.Size([4, 2, 16, 336, 336]) x = (attn @ v) * 0.125 #torch.Size([4, 2, 16, 336, 16]) x = x.transpose(3, 4).reshape(T, B, C, N) x = self.attn_lif(x) x = x.flatten(0, 1) x = self.proj_bn(self.proj_conv(x)).reshape(T, B, C, H, W) #torch.Size([4, 2, 256, 21, 16]) return x class SpikingTransformer(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., norm_layer=nn.LayerNorm, sr_ratio=1): super().__init__() self.norm1 = norm_layer(dim) self.attn = SpikingSelfAttention(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = MLP(in_features=dim, hidden_features=mlp_hidden_dim, drop=drop) def forward(self, x): x = x + self.attn(x) x = x + self.mlp(x) return x class SpikingTokenizer(nn.Module): def __init__(self, img_size_h=128, img_size_w=128, patch_size=4, in_channels=2, embed_dims=256): super().__init__() self.image_size = [img_size_h, img_size_w] patch_size = to_2tuple(patch_size) self.patch_size = patch_size self.C = in_channels self.H, self.W = self.image_size[0] // patch_size[0], self.image_size[1] // patch_size[1] self.num_patches = self.H * self.W self.proj_conv = nn.Conv2d(in_channels, embed_dims//8, kernel_size=3, stride=1, padding=1, bias=False) self.proj_bn = nn.BatchNorm2d(embed_dims//8) self.proj1_lif = MultiStepLIFNode(tau=tau_thr, detach_reset=True) self.maxpool1 = torch.nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False) self.proj1_conv = nn.Conv2d(embed_dims//8, embed_dims//4, kernel_size=3, stride=1, padding=1, bias=False) self.proj1_bn =